Leow Wee Kheng CS4243 Computer Vision and Pattern Recognition. Robust Methods. CS4243 Robust Methods 1

Transcription

1 Leow Wee Kheng CS4243 Computer Vision and Pattern Recognition Robust Methods CS4243 Robust Methods 1

2 Consider this data set Fitting a line to blue points give blue line. Outliers cause fitting error (red line). 25 y outliers x CS4243 Robust Methods 2

3 Consider this data set Fitting a line to blue points give blue line. Single outlier can cause serious error! Need robust methods! y x outlier CS4243 Robust Methods 3

4 Robust Methods Two basic approaches: Make error functions immune to outliers. Detect and discard outliers. CS4243 Robust Methods 4

5 Exercise Consider this set of number: 1, 2, 3, 4, 5, 6, 7 What is the mean? What is the median? Now, insert an outlier 1, 2, 3, 4, 5, 6, 7,14 What is the mean, error of mean? What is the median, error of median? Which one robust, mean or median? CS4243 Robust Methods 5

6 Robust Statistics Mean square error is not robust E = 1 N ( f ( x ) ) 2 Median square error is robust i i y i E = median x i ( f ( ) y ) 2 i i Can potentially tolerate up to N / 2 outliers! CS4243 Robust Methods 6

7 Robust Error Functions Not adversely affected by outliers. Consider least square fitting Want to find parameters a that minimise E. Error is also called residue. Error function can be rewritten as Square of r increases very fast with r. Not robust. CS4243 Robust Methods 7

8 Robust functions ρ(r) increase slowly with r. Properties of ρ(r): Global minimum: ρ(r) = 0 when r = 0 Positive: ρ(r) 0 Symmetric: ρ(r) = ρ( r) Monotonically increasing: ρ(r) ρ(s) iff r s CS4243 Robust Methods 8

9 Robust Error Functions Beaton and Tukey fixed for large r Cauchy increases slowly with r CS4243 Robust Methods 9

10 Robust Error Functions Huber Forsyth increases linearly with r almost fixed for large r CS4243 Robust Methods 10

11 Robust Error Functions CS4243 Robust Methods 11

12 Random Sample Consensus RANSAC: Random Sample Consensus Robust approach for solving many problems. Try to identify outliers. Basic ideas RANSAC wraps over target algorithm. Start with small random subset of data points. Run target algorithm on subset. Iteratively add consistent data points to subset. CS4243 Robust Methods 12

13 Target algorithm has two things Model M: function parameters a, etc. Error measure E: need not be robust Example: least square fit model error Example: affine transformation error model CS4243 Robust Methods 13

14 RANSAC Given N data points. Initialise best model M*, error E(M*). Repeat for k iterations: 1. Randomly select subset S of m < N data points. 2. Run target algorithm on S to determine M. 3. For each data point x not in S If error of M on x < tolerance τ, add x to S. 4. If S > threshold Γ Use S to determine new M. If E(M) < E(M*), update M* M, S* S. consistent set CS4243 Robust Methods 14

15 RANSAC has 3 parameters Error tolerance τ Dependent on the expected error of fitting inliers. Size threshold Γ Dependent on the model. Should be large enough to have enough inliers. Number of iterations k Should be large enough to get good model. Can result in large execution time. CS4243 Robust Methods 15

16 Minimum Subset Random Sampling Use robust error function and random sampling. Basic algorithm 1. Initialise best model M*, error E(M*). 2. Randomly select k subsets of m data points. 3. For each subset S of m data points Determine model M that best fits data points. Compute error E of applying M on all N data points. If E(M) < E(M*), update M* M, S* S. Algo complexity: O(kN) must be immune to outliers CS4243 Robust Methods 16

17 How to be immune to outliers? Use robust statistics Median (r 1,, r N ) Use error function with fixed value for larger r Beaton and Tukey fixed for large r CS4243 Robust Methods 17

18 Use truncated error function ρ( r) = e( r) if r < a e( a) otherwise e(r) is any error function ρ(r) is fixed for large r CS4243 Robust Methods 18

19 Selection of k Max value of k is very large: N C M In practice, smaller value of k is possible. Analysis ε = prob. randomly selected data point is inlier ε = fraction of inliers in data set ε m = prob. all m data points in subset are inliers s = prob. at least 1 subset has only inliers Then, CS4243 Robust Methods 19

20 s = 0.5 s = 0.95 ε = 0.75 k1 k2 ε = 0.5 k3 k4 small k CS4243 Robust Methods 20

21 Example Fit ellipse to data points [RL93] CS4243 Robust Methods 21

22 Summary Outliers can adversely affect algorithm s error. Robust error functions are immune to outliers. RANSAC Robust method for finding consistent set of inliers. Minimum subset random sampling Use robust error function and random sampling to speed out random sampling. CS4243 Robust Methods 22

23 Further Reading Robust parameter estimation [Ste99], [MMRK91] RANSAC [FB81], [FP03] Section 15.5, Minimal Subset Random Sampling [RL93] CS4243 Robust Methods 23

24 References M. A. Fischler and R. C. Bolles. Random sample consensus: A paradigm for model fitting with apphcatlons to image analysis and automated cartography. Communications of ACM, 24(6): , D. A. Forsyth and J. Ponce. Computer Vision: A Modern Approach. Pearson Education, P. Meer, D. Mintz, A. Rosenfeld, and D. Kim. Robust regression methods in computer vision: A review. Int. Journal of Computer Vision, 6(1):59 70, G. Roth and M. D. Levine. Extracting geometric primitives. CVGIP: Image Understanding, 58(1):1 22, C. V. Stewart. Robust parameter estimation in computer vision. SIAM Review, 41(3): , CS4243 Robust Methods 24